The Big Picture
Personalizing an agent’s emotion perception during a live negotiation makes people concede more and lets the agent secure better deals without harming overall fairness.
ON THIS PAGE
The Evidence
Continual personalization of facial emotion signals (updating a perception model during interaction) produced a stronger, more varied arousal signal that the agent used to adjust its offers. With the same negotiation policy, the personalized system caused people to concede more and increased the agent’s utility versus a fixed, generic emotion model. These gains came with no significant drop in joint outcome quality (distance to the Nash solution) and were achieved using an asynchronous setup Orchestrator-Worker Pattern that learns in the background so interaction stays responsive.
Not sure where to start?Get personalized recommendations
Data Highlights
1Agent utility increased (paired test): p = 0.017, effect size d = 0.30.
2Participant concession rate rose by 9.35 percentage points (p = 0.013, d = 0.31).
3Personalization raised estimated arousal (subject-level effect d = 0.68) for about 76% of participants.
What This Means
Engineers building human-facing negotiation agents or sales bots can use on-the-fly personalization to improve outcomes without changing core strategy. Product and research leads evaluating agent reliability or user-facing responsiveness will find this approach useful for boosting persuasiveness and perceived awareness while preserving fairness.
Key Figures
![Fig. 1: Stepwise Move Classification [ 24 ] : Illustrated move types are defined by both parties’ utility changes from the same actor’s previous offer: Fortunate (both gain), Nice (opponent gains, own unchanged), Concession (own loss, opponent gain), Selfish (own gain, opponent loss), Unfortunate (both lose).](https://arxiv.org/html/2611.14247v1/2611.14247v1/figures/MoveAnalysis.png)
Fig 1: Fig. 1: Stepwise Move Classification [ 24 ] : Illustrated move types are defined by both parties’ utility changes from the same actor’s previous offer: Fortunate (both gain), Nice (opponent gains, own unchanged), Concession (own loss, opponent gain), Selfish (own gain, opponent loss), Unfortunate (both lose).

Fig 2: Fig. 2: System Architecture of the CL-enabled Solver Agent: The robot’s audiovisual input is converted to offers (Speech-to-Offer) and valence–arousal estimates (CLIFER: episodic memory adapted to the participant, CAAE providing generative replay). The Solver uses these to compute counter-offers, which the Interaction Manager voices with gestures. θ \theta denotes the GWR memory parameters; the FaceChannel encoder and regression head remain fixed.

Fig 3: Fig. 3: Differential affective feedback between the previous ( E ~ t − 1 \tilde{E}_{t-1} , blue) and current ( E ~ t \tilde{E}_{t} , orange) normalized VA vectors. Equation 7 defines angular ( P r o t P_{rot} ) and radial ( P r a d P_{rad} ) features; the dashed line shows displacement ( Δ E ~ \Delta\tilde{E} ).

Fig 4: (a) Participant view
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreConsiderations
The system relied primarily on facial signals, so performance may change in noisy, low-visibility, or voice-heavy settings. A short demo session was used to warm up the personalized model, which may not be available in all real-world deployments. The study did not independently verify ground-truth emotion labels, nor did it measure long-term retention across sessions or catastrophic forgetting in longitudinal use. Context Drift as a potential failure mode could affect reliability in changing environments.
Methodology & More
A negotiation agent that adapts its facial emotion perception during interaction gains practical advantages. The approach combines a continuous two-dimensional emotion estimate (valence and arousal) with a simple control rule that turns directional and intensity changes in that estimate into modest shifts in the agent’s target offer utility. The personalized perception model is updated continuously using a memory- and replay-based routine, while inference and offer selection run on a separate, fast thread so the interaction stays under real-time constraints.
In a within-subject user study with 66 participants (each experienced both the fixed baseline and the adaptive system), personalization produced a wider and stronger arousal signal, increased participant concession rates by 9.35 percentage points, and raised agent utility with a small-to-moderate effect size. Joint optimality (measured as distance to the Nash bargaining point) and participant utility did not worsen. The work shows a practical path to make socially aware agents more effective: adapt perception to the person in front of you, feed that signal continuously into strategy, and keep learning asynchronous so conversations feel natural. Next steps would add voice and text signals, test longer-term retention, and evaluate performance outside controlled demo conditions. Semantic Capability Matching Pattern and Dynamic Task Routing Pattern could further support robust, personalized negotiation behavior.
Avoid common pitfallsLearn what failures to watch for
Credibility Assessment:
Authors include mid-career researchers with moderate h-indices (≈16–18) and a university affiliation (Ozyegin) — recognized but not top-tier signals, so a solid/average credibility.